DataPDF Available

MSLDA14 Keynote

Data

MSLDA14 Keynote

Figures

No caption available
… 
No caption available
… 
Content may be subject to copyright.
!"#$%$&$"'(')'#*&'((+(,-.-#!/(01$,"2#-3-#)(
,45(1676(894:6; ( ((((
!"#$%#&'()*+,+'"%-).%#+$/,0 ) ))
*+,+'"%-)1"23&)4+'5+")
6#,0"#730+5).+$,#$8).9,0+:,) ) (((()
/<=>?@A4B(@46C;D=>(D=(E6@B<=(
/AFF<=F(6=:($GA=(2H6IIA=>AF(
(
&/',!(JKLM(
/<=>?@A4B#46C;D=>(<N('B6II($O7AC@F(
;-9)#,)#0)#:&2"0'$0)<2")=3,0"'(#'>)
?-'((+$8+,@),&'",+)&2&3('/2$A)('"8+)('$5:',,)
=&&(#%'/2$,@)'8"#%3(03"+A)7#2,+%3"#09A)(28#,/%,)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
I))F)
#46C;D=>(P(*QDFR=>('<I9R<=F(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
4#<+/:+)
.':&(#$8)J"+K3+$%9)
.-2"0B0+":)
<"+K3+$0)
,':&(#$8)
42$8B0+":)
,&'",+),':&(#$8)
5'9,) :2$0-,) 9+'",)
L))F)
#46C;D=>(P($94(.<6I(
.':&(#$8)J"+K3+$%9)
.-2"0B0+":)
<"+K3+$0)
,':&(#$8)
42$8B0+":)
<"+K3+$0)
,':&(#$8)
42$8B0+":)
,&'",+),':&(#$8)
4#<+/:+)
5'9,) :2$0-,) 9+'",)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
M))F)
26FA('@9:ST(#46C;D=>(UISD=>(U<QAF(
6#,+',+)
.&"+'5)
N O+$5"')P#"3,)
N Q72(')#$)=,#'R=<"#%')
N ?2"2$'S#"3,)#$)!+",#'$)13(<)
.++5)
6#,&+",'()
N D#2).+%3"#09)
D+-'S#23")
N T-"+'0+$+5),&+%#+,)
N U$0+"'%/2$)V#0-)20-+")'$#:'(,)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
W))F)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
X))F)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
Y))F)
UISD=>(U<QAF(D=(!9F@46ID6T(VH6@(:<(WA(;=<WX(
MZZ[)C$2V$)\9#$8)<2])%':&,)',,+,,+5)+S+"9)L):2$0-,)79):'$3'()
,3"S+9,)
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
^))F)
0I6Y<4B(1AZ9D4ABA=@F(
6+S+(2&:+$0)2<)'):27#(+)0"'%C#$8)&('_2":)<2")\9#$8)<2]+,@)
4#8-0BV+#8-0)`aIZ)8"':,b)
!"2(2$8+5)5+S#%+)(#<+/:+)`:2$0-,R9+'",b)
O#8-)"+,2(3/2$)&2,#/2$)5'0')
?2$0+]0)`<++5#$8A)\9#$8A),(++&#$8A)5+<+%'/$8b)
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
c))F)
[64:W64A(6=:('<\W64A(!4CHD@AC@94A(
26B6]<@](#46C;D=>('SF@AB(
(
?':'d20d@)D'0)825)#$)e'9'$):90-2(289)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
fZ))F)
26B6]<@](&<ODIA(#46C;D=>(0I6Y<4B(
e3(/B:25'(),+$,#$8)&('_2":)V#0-),-2"0B"'$8+)"'5#2)0"'$,%+#S+"@)
N1!.)"+%+#S+")
NU$+"/'(),+$,2",)`'%%+(+"2:+0+"A):'8$+02:+0+"b)
N!"+,,3"+)'$5)':7#+$0)0+:&+"'03"+),+$,2")
Ne#%"2&-2$+)
)
)
*g)H3"5'CA)!g).2::+"A)Dg)h3,9A)+0)'(g)ie3(/:25'()=%/S#09B7',+5)1!.).':&(#$8Aj)-0'%)IZfLg)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
ff))F)
,AI6S?#<IA46=@(VD4AIAFF(%A@W<4;D=>(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
fI))F)
,AI6S?#<IA46=@(VD4AIAFF(%A@W<4;D=>(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
fL))F)
,AI6S?#<IA46=@(VD4AIAFF(%A@W<4;D=>(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
fM))F)
%A@W<4;(<N(E6FA('@6R<=F(
fZZ)C:)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
fW))F)
,AI6S?#<IA46=@(%A@W<4;D=>(
U$5#S#53'(,)0"'S+()7+0V++$)5#k+"+$0)%':&,)'$5)20-+")(2%'/2$,)
.02"+),+$,2"R&2,#/2$),':&(+,)(2%'((9)#$)\',-)
l&(2'5)3,#$8),-2"0B"'$8+)"'5#2)02)8'0+V'9)`L1b)'0)C$2V$)%':&,)
)
D',+)=)
D',+)D)
D',+)?)
6D)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
fX))F)
2944A=@(@46C;A4F(W<4;(WAII55(O9@(=<@(N<4(I<=>(
?'$)V+)270'#$)0-#,)5'0')K3'(#09)$+'"B&+"&+03'((9>)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
fY))F)
^AS(2H6IIA=>AF(
Q$+"89)
)
e27#(#09)
6+('9B02(+"'$%+)
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
f^))F)
*=A4>S(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
fc))F)
26=(WA(C<=R=9<9FIS(@46C;(G<FDR<=(WD@H(O<9=:A:(
9=CA4@6D=@SX(
(
1!.)
.%-+53(+")
Ig)1!.)6309)
?9%(#$8)
Lg)=%/S#09)
6+0+%/2$)
Mg)e27#(#09)
e25+(,)
1!.).':&(#$8)
/:+,)'$5)
J"+K3+$%#+,)
fg)Q$+"89)
Q,/:'/2$)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
IZ))F)
$=ID=A(*=A4>S(*FRB6R<=(
O2V)02)+,/:'0+)7'm+"9),0'0+)2<)%-'"8+>)
D'm+"9)S2(0'8+)3$"+(#'7(+)#$):#5B"'$8+A)9+0)"+('/S+(9)'%%3"'0+)
#$)+]0"+:+),0'0+,)
.2nV'"+):+0+"#$8)%'&03"+,)#$,0'$0'$+23,)$+0)+$+"89)\2V)9+0)
#,),3,%+&/7(+)02)(2$8)0+":)5"#n,)
1.0
0.9 0.8
0.7
0.6 0.5
0.4
0.3
0.2 0.1
0.0
State of Charge (SOC)
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
4.0
4.1
Battery Voltage [V]
29
62
99
119
240
468
1000
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
If))F)
42$8B0+":)5"#n)
U$,0'$0'$+23,).E?)
S'"#'$%+)
D'm+"9)
P2(0'8+)
.2nV'"+)
D22CC++&#$8)
?2$\'/2$B
7',+5)
'&&"2'%-)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
II))F)
'$2(AFRB6R<=(@H4 <9>H(C <=_ 6R<=(
Figure 1. Power consumption for different tasks: Baseline consumption is roughly 230µA (left), inertial task running for
2 seconds at t=0s and radio enabled for 1 second at t=30 (center), and GPS hotstart with 30 seconds of tracking (right).
The scheduler keeps track of the total runtime for each
task and can thus estimate the energy consumption of the
system if the current drawn by each sensor is known. This
can be measured prior to the deployment (see Table 1).
4.3 Measuring Battery Voltage
Battery voltage is commonly used to estimate SOC of
the battery. Figure 2 shows discharge curves of our battery
(300 mAh Tenergy Lithium-Ion Polymer 3.7 V) for resistive
loads between 29 and 1000 Ohm. We measure the time in-
terval from a fully charged state until the voltage falls below
3.1 V and calculate the battery SOC at any point during this
interval.
Battery voltage provides a good estimation of SOC if the
battery is nearly fully charged or empty. However, the slope
in the discharge curve is relatively flat over large parts of the
battery capacity and depends on the instantaneous discharge
current. Thus, the voltage-based SOC estimation has limited
accuracy in this middle region, due to the voltage jitter under
different operating conditions.
Figure 2. State of charge when discharging using differ-
ent load resistances.
5 Estimation of the State of Charge
We use a closed-loop model to estimate the battery SOC
for each time interval, as shown in Figure 3. We define the
estimation of the battery SOC at the discrete time instant t(k)
as a value between 0 (battery empty) and 1 (battery full):
d
SOC[k] 2 [0, 1] (4)
The estimation of the state of charge is updated in every
time step as the weighted sum of two separate estimators for
SOC based on energy (
d
SOC
e
) and battery voltage (
d
SOC
v
):
d
SOC[k]=w
e
[k] ·
d
SOC
e
[k 1]+w
v
[k] ·
d
SOC
v
[k 1] (5)
d
SOC
Energy Estimation
d
SOC
v
d
SOC
e
Voltage Measurement
w
e
w
v
DE
Figure 3. Model for estimation of state of charge (SOC).
The weighting factors w
e
and w
v
determine the influence
of each separate estimator on the final estimation of
d
SOC
in each time step. One important difference between the
voltage- and energy-based SOC estimators is that voltage-
based estimator does not depend on a previous SOC estimate,
while the energy-based estimator does. However, deriving
the SOC from the battery voltage has a larger uncertainty
when the battery is not close to the fully charged or empty
state (see Figure 2).
Instead of representing SOC estimators as a single num-
ber, we represent them as normal distributions given by their
mean m and variance s
2
. We use a mathematical method
called conflation [7] to combine data from two noisy sources
estimating the same quantity (SOC) and calculate the new
normal distribution. Conceptually, the distribution with the
smaller uncertainty has a larger influence on the new mean.
The new mean m and variance s
2
are calculated as follows:
m =
s
2
v
· m
e
+ s
2
e
· m
v
s
2
e
+ s
2
v
, s
2
=
s
2
e
· s
2
v
s
2
e
+ s
2
v
(6)
Therefore, we can derive the weighting factors w
e
and w
v
accordingly:
w
e
=
s
2
v
s
2
e
+ s
2
v
, w
v
=
s
2
e
s
2
e
+ s
2
v
(7)
6 Experimental Results
In this section, we evaluate the methods proposed in this
paper by means of controlled experiments and using empiri-
cal data from animal collars collected over several weeks.
6.1 Accuracy of Power Profiling
In order to evaluate the accuracy of our power profiling
method proposed in Section 4, we performed a lab experi-
ment for the duration of 12 hours. During that interval, we
configure the sensor node to log GPS position fixes every 12
minutes for a period of 30 seconds. Furthermore, we sample
!g).2::+"A)Dg)h3,9A)'$5)*g)H3"5'CAoQ$+"89)Q,/:'/2$)<2")42$8B0+":)T"'%C#$8)
=&&(#%'/2$,AoA)`*='SFb)IZfLg)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
IL))F)
0
1 2
3
4
5
6
7
8 9
10 11 12 13 14 15 16 17 18 19 20 21 22 23
Day
0.0
0.5
1.0
0
5
10
15
[mA]
3.2
3.6
4.0
4.4
Voltage
'$2(*FRB6R<=(1AF9I@F(
Figure 4. Estimated average current consumption for GPS and data download tasks (top: Node A, bottom: Node B).
Figure 5. Estimation of battery SOC based on sensor data from free-living flying foxes (top: Node A, bottom: Node B).
the inertial sensors every 10 minutes for 2 seconds. Current
consumption by the sensor node is measured at 25 Hz using
a Fluke 8845A precision multimeter. No solar panels were
connected to the sensor node during the experiment. We
observe that the node’s average current draw as measured
by the multimeter is 1.74 mA. In comparison, our software
bookkeeping estimated an average current draw of 1.63 mA,
which is a difference of less than 10%. Thus, we use an un-
certainty value of 0.1 for DE in our model.
6.2 Deployment Data
We collected sensor readings from two sensor nodes de-
ployed on free-living flying foxes. Both animals were col-
lared nearby the same roosting camp on the same day. For
both animals, we take into account the same observation pe-
riod, which spans 23 consecutive days.
SOC Estimation. Based on the sensor data available from
the nodes, we estimate the current consumption when run-
ning different tasks. Our application on the bat logged the
battery voltage, solar charge current, and solar voltage every
15 minutes to external flash. Based on this data, we esti-
mate the amount of harvested solar energy. We also calcu-
late the estimated battery state of charge (
d
SOC) according to
the method proposed in Section 5. Thereby, the initial value
for
d
SOC is set to 50% capacity. We use empirical data from
battery discharge measurements (see Figure 2) to estimate
d
SOC
v
. We model the the uncertainty s
2
v
of
d
SOC
v
with a
value of 0.1 when the battery is nearly full (
d
SOC
v
> 0.9) or
nearly flat (
d
SOC
v
< 0.1), while s
2
v
is set to 0.5 otherwise.
Results. We plot the current consumption of the GPS and
data download tasks averaged over 1-hour intervals in Fig-
ure 4. Data is downloaded from Node A almost daily when
the animal returned to the roosting camp, while Node B
downloaded data in bulk when returning to the camp for the
first time after 20 days.
Figure 5 shows the measured charge current (I
in
), the soft-
ware estimated current draw by the system (I
out
), and the
resulting delta in terms of energy over time (DE). Node A
starts with a battery which is almost empty as indicated by
the large variations in the battery voltage. Consequently, the
voltage measurement has larger weight in the estimation of
SOC. Although the energy net flow into the battery is posi-
tive, it takes roughly 10 days until the battery voltage stabi-
lizes which corresponds to an estimated state of charge above
0
1 2
3
4
5
6
7
8 9
10 11 12 13 14 15 16 17 18 19 20 21 22 23
Day
0
1
2
3
4
5
Current [mA]
GPS
Data Download
!g).2::+"A)Dg)h3,9A)'$5)*g)H3"5'CAoQ$+"89)Q,/:'/2$)<2")42$8B0+":)T"'%C#$8)
=&&(#%'/2$,AoA)`*='SFb)IZfLg)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
IM))F)
.0'(,9@S(2SCID=>(
1!.)p])
X Assumed position
Real position
Uncertainty
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
IW))F)
.0'(,9@S(2SCID=>(
!2V+")1!.)7'%C)2$)V-+$)3$%+"0'#$09)+,/:'0+)'&&"2'%-+,)723$5)
1!.)2k)
X Assumed position
Real position
Uncertainty
X
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
IX))F)
GPS Duty Cycling Strategy
Varying the AAU according to
the animal’s distance from the
fence
Speed models
AAU: absolute acceptable uncertainty
U
gps
: GPS chip uncertainty
s: assumed speed
t
L
: lock time
*g)H3"5'CA)!g)?2"C+)+0)'(gA)oQ$+"89B+q%#+$0)42%'(#,'/2$@)1!.)6309)?9%(#$8)V#0-)*'5#2)*'$8#$8Ao))=?e)TE.rA)P2(g)cA)U,,g)IA)e'9)IZfLg))
*g)H3"5'CA)!g)?2"C+A))+0)'(g)o=5'&/S+)1!.)6309)?9%(#$8)'$5)*'5#2)*'$8#$8)<2")Q$+"89BQq%#+$0)42%'(#d'/2$Aj).+$,9,)IZfZg))
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
IY))F)
Exploiting Radio Proximity Data
Animals naturally herd closely
together
GPS duty cycling vs GPS
DC and contact logging
Combining GPS duty cycling
with short range radio beaconing
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
I^))F)
.0'(,9@S(2SCID=>(WD@H(2<=@6C@(/<>>D=>(
l,#$8)$+#8-72",)',)&2,#/2$)'$%-2",)02)723$5)
3$%+"0'#$09)
)
1!.)2k)
X Assumed position
Real position
Uncertainty
X
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
Ic))F)
.0'(,9@S(2SCID=>(WD@H(2<=@6C@(/<>>D=>(
l,#$8)$+#8-72",)',)&2,#/2$)'$%-2",)
)
X Assumed position
Real position
Uncertainty
X
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
LZ))F)
'A=F<4?@4D>>A4A :(. 0 '('6BGID=>(
Activity
Sensors Timing
Audio Inertial
Air
Solar
Event Event GPS Sampling
Pressure Duration Frequency Period
Flying X X hours daily high
Interacting X X seconds frequent on event
Urinating/Defecating X seconds frequent on event
Grooming X X seconds very frequent none
Resting X X X hours daily infrequent
Table 4: Key activities of ying foxes, their timing prole, and the sensors we use to detect them.
















Time (s)
          







Relative Sound Level (dB)









Frequency (kHz)
        
Time (s)

Sound Level (dB)
        
Time (s)

Normalized frequency
Mean
sound level
Duration of
sound event
Mean
normalized
frequency
Figure 8: Top: spectrogram of typical audio interac-
tion event. Middle: corresponding sound level and
zero crossings. Bottom: normalised frequency. Ar-
rows show derived acoustic features.
ity classification. Figure 9 plots accuracy, precision and the
performance metric [15] as the threshold is increased from
0. The performance metric is the product of accuracy, preci-
sion, sensitivity and specificity and serves as an indicator for
selecting a threshold which gives the highest accuracy while
maintaining a high level of precision. Figure 9 also shows the
receiver operator characteristic (ROC) curve which plots the
true positive rate vs. the false positive rate as the threshold
is varied. The indicated operating point corresponds to the
selected threshold of 0.002. Two-fold cross validation was
done over 1000 iterations to evaluate the performance of the
classification by splitting the dataset in half. This resulted
in a mean accuracy of 77.5 % and a mean precision of 70.5 %
relative to manually marked ground truth obtained via video
footage and external audio recordings.
3.4.2 Inertial
The inertial sensors on our platform enable the detection
of activities such as interaction among multiple animals, uri-
nating/defecating, and grooming behaviour, either individ-
ually or in combination with other sensors. For instance, ac-
celerometers can be combined with acoustic sensor data to
Threshold [x10
-2
]
0.0
0.2
0.4
0.6
0.8
1.0
Accuracy, Precision & Performance
0.0 1.0 2.0 3.0 4.0
Accuracy
Performance metric
Precision
Used threshold
False positive rate
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
True positive rate
Operating point
Figure 9: Plot of accuracy, precision and perfor-
mance metric vs. classification threshold (left), and
receiver operator characteristic (ROC) curve for
acoustic activity classification (right) showing the
operating point corresponding to the used thresh-
old of 0.002.
detect interactions among multiple animals. Alternatively,
accelerometers can independently detect the full reversal of
orientation that occurs when flying foxes engage in waste
removal from their bodies.
We examine accelerometer signals collected from a fly-
ing fox collar at 128 Hz during the captive bat experiments.
Video footage and visual inspection serve as the ground
truth for this experiment. In order to visually distinguish
the angular inversion that occurs during urination activi-
ties, we compute the mean three-dimensional vector during
a 7 min portion of the experiment. The reason for choosing
the me an vector is that the flying fox remains in a down
facing position for most of the experiment, which indicates
that the mean vector should provide a decent estimate of
the constant gravitational force and serve as a reference for
orientation reversal. Figure 10 (top) shows the XYZ com-
ponents of the accelerometer signal projected on the mean
vector. There are clear sign inversions in all the accelerome-
ter dimensions in two instances in the trace. However, using
sign inversions to detect orientational flips is susceptible to
corner cases where one of the accelerometer dimensions i s
orthogonal to the gravity vector.
We detect inversion events instead by computing the an-
gle between the current 3-D acceleration vector ~c and the
inferred gravity vector ~g, using the following equation:
tan()=norm (~g ~c, ~g · ~c)(1)
where is in degrees, and norm i s the vector norm function.
The rationale for using angular shifts is that any 180° inver-
sion in orientation will result in a significant shift in that
is greater than 90° for a sustained period, which can only
correspond to waste removal events in flying foxes.
Figure 10 (bottom) shows the resulting angles. The rect-
angular boxes indicate two detected instances of inversion
events, while the left and ri ght i mages show the correspond-
l,+)2$+)2"):2"+)2<)0-+)%-+'&)2$B72'"5),+$,2",)
02)5+0+%0)'%/S#/+,)2<)#$0+"+,0)'$5)0"#88+")1!.)
,':&(+,))
.2:+)'%/S#/+,)2<)#$0+"+,0)
*g)H3"5'CA)!g).2::+"A)Dg)h3,9A)+0)'(g)ie3(/:25'()=%/S#09B7',+5)1!.).':&(#$8Aj)-0'%)IZfLg)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
Lf))F)
'A=F<4?@4D>>A4A:(.0'(F6BGIAF(`!CCAIA4<BA@A4a(
?2:&30+)'S+"'8+)
S+%02")'0)"+,0)!)8"'S#09)
?2:&30+)'$8(+)
7+0V++$)%3""+$0)S+%02")
'$5)8"'S#09)
6+0+%0),3,0'#$+5)
'$83('"),-#n,)'72S+)
cZ
2)
fZZs)'%%3"'%9)#$)
5+0+%/$8)ff)0"3+)
+S+$0,)
P#5+2)<220'8+)',)
8"23$5)0"30-)
)
1.4 1.5 1.6 1.7 1.8 1.9 2
x 10
5
0
100
200
Sample
Angle current and
gravity (degrees)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
LI))F)
'A=F<4?@4D>>A4A:(.0'(F6BGIAF(`!9:D<a(
Inertial
















Time (s)
          







Relative Sound Level (dB)









Frequency (kHz)
        
Time (s)

Sound Level (dB)
        
Time (s)

Normalized frequency
Mean
sound level
Duration of
sound event
Mean
normalized
frequency
J"+K3+$%9)&+'C,)'0)IBMh-d))
4#8-0V+#8-0)<+'03"+,)'"+)7',+5)2$)
%'(%3('/$8)0-+):+'$),#8$'()+$+"89)'$5)
%23$/$8)0-+)$3:7+")2<)d+"2)%"2,,#$8,)
2<)')fZIM),':&(+),(#5#$8)V#$52V)V#0-)
'$)2S+"('&)2<)WZs)
P#5+2)< 220'8+)',)8"23$5 )0"30-)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
LL))F)
'A=F<4?@4D>>A4A:(.0'(F6BGIAF(`!9:D<a(
Inertial
















Time (s)
          







Relative Sound Level (dB)









Frequency (kHz)
        
Time (s)

Sound Level (dB)
        
Time (s)

Normalized frequency
Mean
sound level
Duration of
sound event
Mean
normalized
frequency
Threshold [x10
-2
]
0.0
0.2
0.4
0.6
0.8
1.0
Accuracy, Precision & Performance
0.0 1.0 2.0 3.0 4.0
Accuracy
Performance metric
Precision
Used threshold
False positive rate
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
True positive rate
Operating point
J"+K3+$%9)&+'C,)'0)IBMh-d))
4#8-0V+#8-0)<+'03"+,)'"+)7',+5)2$)
%'(%3('/$8)0-+):+'$),#8$'()+$+"89)'$5)
%23$/$8)0-+)$3:7+")2<)d+"2)%"2,,#$8,)
2<)')fZIM),':&(+),(#5#$8)V#$52V)V#0-)
'$)2S+"('&)2<)WZs)
P#5+2)< 220'8+)',)8"23$5 )0"30-)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
LM))F)
&9IRB<:6I(*bA=@(:DFF<CD6R<=(
;-+$)2$+),+$,2") #,)
#$,3q%#+$0)02)%'&03"+)
+S+$0B2<B#$0+"+,0)
Q]':&(+@)-2V)02)
5#,,2%#'0+)#$0+"'%/2$)
+S+$0,)#$S2(S#$8)')
%2(('"+5)'$#:'()<"2:)
#$0+"'%/2$)+S+$0,)
#$S2(S#$8)$+'"79)
'$#:'(,)2$(9>)
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
LW))F)
&9IRB<:6I(*bA=@(:DFF<CD6R<=(
;-+$)2$+),+$,2") #,)
#$,3q%#+$0)02)%'&03"+)
+S+$0B2<B#$0+"+,0)
Q]':&(+@)-2V)02)
5#,,2%#'0+)#$0+"'%/2$)
+S+$0,)#$S2(S#$8)')
%2(('"+5)'$#:'()<"2:)
#$0+"'%/2$)+S+$0,)
#$S2(S#$8)$+'"79)
'$#:'(,)2$(9>)
)
0 400 800 1200
2
0
2
Time (seconds)
Acceleration
ACC
X
ACC
Y
ACC
Z
Detected Interaction Events
0 400 800 1200
0
50
100
150
Time (sec)
Angle (degrees)
Changes in mean
angular shift
Angular shift
Time (s)
400 800 12000
Mean sound level (dB)




0
200 600 1000
Acoustic activity
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
LX))F)
&9IRB<:6I(!CRbD@S?O6FA:(/<C6IDF6R<=(
Collared events Nearby events Power consumption
Detected Events
Average Power Consumption (mW)
Accelerometer MAL
collared only
MAL
nearby only
MAL
all events
4.5
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
9
8
7
6
5
4
2
0
1
3
Audio
Localisation Approach
Animal interactions
Collared All Dissociated
Duty cycled GPS X
Accelerometer-triggered X
Audio-triggered X
Accel. AND Audio X
Accel. OR Audio X X
Table 5: MAL can detect all events and disso ciate
interaction event involving collared animal or nearby
animals.
in our simulations. We compare a baseline approach of a
duty cycled GPS with a period of 20 s with triggered GPS
sampling approaches based on the accelerometer only, audio
only, or on the combination of audio and accelerometer sen-
sors. We group all detected ground truth interactions into
events that meet the 25 s to 1 min duration constraint. A
successful detection in our simulation is when the algorithm
obtains at least one GPS sample during the event.
During the given time window, the duty cycled GPS mod-
ule remains active for a total of 451 s (including lock times)
and successfully obtains GPS samples during each of the
four events of interest, yielding an overall node power con-
sumption of around 33 mW. Figure 13 summarises the re-
sults of sensor-triggered GPS sampling. The accelerometer-
triggered GPS manages to detect only two events (only the
events from the collared bat) with a cumulative GPS active
time of 21 s and power saving of 86 % over the GPS duty cy-
cled approach. In comparison, the audio-triggered GPS can
detect all four interaction events of interest while keeping the
GPS active for a total of 64 s, corresponding to a node power
consumption of 7.42 mW. However, the audio-triggered ap-
proach can only determine that interaction events are occur-
ring nearby, but not w hether the collared animal is involved.
MAL can be tuned to capture only interaction events in-
volving the collared animal, with comparable detection to
accelerometer and s lightly higher power consumption for
powering the audio sensor. Alternatively, MAL can be tuned
to capture only nearby interaction events, yielding a 14%
reduction in power consumption over audio and correct de-
tection of the two interaction events involving only nearby
animals. Triggering the GPS on the basis of both the ac-
celerometer and audio activity detectors yields comparable
energy consumption to audio and correctly dissociates the
two types of detected events.
The main benefit of MAL is that it provides users with the
flexibility to tune performance to their current activities of
interest. If users are interested in collared bat interactions
only, they can simply use accelerometer triggers for obtain-
ing GPS samples and save energy in the process. If they
are interested in the cumulative set of interaction events
regardless of individual animal association with activities,
then audio is sucient. If, on the other hand, users are in-
terested in pinpointing individual animals associated with
each activity, multimodal triggering of the GPS can provide
the data granularity for dissociating these event types.
5. RELATED WORK
The Networked Cow project [8] used PDAs with GPS and
adhoc-mode WiFi to route position information to a base
station. The work in [6] extends this cattle tracking ap-
plication to use short-range radio for relative localisation
Figure 13: Performance of MAL against
accelerometer- and audio-triggered GPS. MAL
can be tuned to capture either interaction events
of the collared animal, or nearby interaction events
only. MAL can also detect and dissociate both
types of interaction events with comparable power
consumption to audio.
alongside G PS. The ZebraNet project [5] reports i ndividual
position records for zebras every few minutes. In order to
make the energy problem more tractable ZebraNet collars
include a solar panel, which assume that the panels are re-
silient to normal animal activities. Positioning is done by
GPS only, and the nodes propagate their information by
flooding in order to facilitate data acquisition by the mobile
sink. Dyo e al. [3] use a heterogeneous sensor network con-
sisting of RFID-based tags and base stations to track Euro-
pean Badgers over a prolonged period of time and highlight
the importance of interaction with domain scientists and
early prototyping, w hich are also central to our methodol-
ogy in design Camazotz. Our work shares the long-term
monitoring goals and network topology with [3], but Cama-
zotz includes GPS modules on the wildlife tags and aims to
push the size, weight, and lifetime of the nodes to new limits
through aggressive duty cycling based on MAL.
Anthony et al. [1] developed the CraneTracker system for
long-range long-duration tracking of the endangered whoop-
ing Crane. Their platform, weighing about 100 g, includes
GPS and inertial sensors as well as cellular and an Atmel
RF230 radio for short-range communication. Their design
aims at two GPS fixes/day and a communication latency of
less than 24 hours. While our work also targets long-range
and long-duration tracking of small birds, our target applica-
tion tracking flying foxes has much stricter design goals. For
instance, the device can not weigh more than 30 to 50 g or
5% of the bodyweight of the animals. Additionally, we aim
for position logs at the frequency of at least once every half
hour which results in a much higher utilisation of the GPS
module. The combined smaller footprint and higher GPS
sampling frequency for our application motivates our design
of the Camazotz platform. The use of accelerometers has
also been proposed as a low power indicator of movement to
supplement GPS duty cycling [20] [14]. Guo et al. [12] also
consider the use of directional and angular speed for cat-
tle behaviour classification. The work in [7] addresses the
tradeo between localisation accuracy and energy eciency.
Akeydierence with our work is that we use multiple sensor
modalities to trigger GPS duty cycling for more fine-grained
activity detection.
Recently, Liu et al. [16] proposed a sample-and-process
approach to dramatically reduce the active time for GPS
position sampling by up to three orders of magnitude. While
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
LY))F)
2<=c>946R<=(<N('A=FD=>(#6F;F(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
.(20)f)
.(20)I)
.(20)L)
.%-+53(+")
*+'(BT#:+)
?(2%C)
Q$+"89)
e2$#02"#$8)
?2$0+]0)
`:2/2$R$2B:2/2$b)
1!.))
*'5#2))
U$+"/'()
!2V+")
T+:&+"'03"+)
!"+,,3"+)
#6F;(2<=c>946R<=(
'A=F<4(04<CAFF(
)
.(20)$)
T#:+@)f^@ZZ)t)
ZX@ZZ)R)fW):#$)
fZZ)1!.)
,':&(+,uf)Od)
v)
L^))F)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
&<ODID@S(
Lc))F)
&<ODID@S(&<:AID=>(
Q,0'7(#,-):27#(#09)5+&+$5+$%#+,)
T+:&2"'()
.&'/'()
Q$S#"2$:+$0'()
.2%#'()
?-'"'%0+"#d+)&2&3('/2$)'$5)#$5#S#53'()(+S+():2S+:+$0)
,0'/,/%,)
.0+&),#d+)
T3"$#$8)'$8(+)
6#k3,#2$))
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
MZ))F)
*QGI<46R<=d*QGI<D@6R<=(U<46>D=>(&<:AI(
E7,+"S+5)&'m+"$,)2<)5#,&('%+:+$0)5#,0"#73/2$)<2"):'$9)
,&+%#+,)'&&"2]#:'0+5)79)')&2V+")('V)2<)0-+)<2":@)
;-+"+)fawaL)5+0+":#$+,)0-+)09&+)2<):2/2$)"'$8#$8)<"2:)7'((#,/%)
02)D"2V$#'$)
)
T9&#%'()'&&"2'%-)02)8+0)w)<2")'),&+%#+,)#,)0-"238-)+:&#"#%'()5'0')
2")',,3:&/2$,)
O2V)52+,)0-+)2&/:'()w)"+,&2$5)02)0-+):2S+:+$0)<+'03"+,)2<)
0-+)0"'G+%02"9>)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
For Review Only
1 Introduction
Understandin g anima l movement is crucial for understan d i n g ecologi ca l and evolutionary
processes in nature a n d has a wide range of applica t i o n s such as ecosystem management,
speci es conservation and disease control [1, 2, 3, 4, 5].
In the last decad e it has been widel y observed that t h e movement s of many animal
speci es, from albatrosses [6, 7 ] to spider monkeys [8], honey bees [ 9 ] to deer [10 ], and
marine predators [11] to hum a n forager s [12, 13], appear to exhibit L´evy-flight patterns,
i.e. the step -si ze distri b u t i o n can be approximated by a power-law P (l) l
µ
with
1 3. Despite this apparent similarity, there is an ongoing debate in the scientific
commu n i ty over the existence of L´evy-flights i n animal movement and the method ology
of verifying L´evy-flights from empirical data [15, 16, 17]. In th e meanwhile, scientists,
especi al l y theorists, are keen on a questi on from a theoretical perspect i ve: if the existen ce
of L´evy-flight in animal movement is true, why do animals perform L´evy-flight? This
question fascinates researchers from various disciplines from ecology to physics [18, 19,
20, 21, 22, 23].
One common approach to the origin of animal movement patterns is to use the scheme
of optimizing random search [24, 25, 26]. In a random search model, singl e or multiple
individua l s search a landscape to locate targets whose locations are not known a priori,
which is usually adopted to describe the scenario of anim al s foraging for food, p r ey
or resources. The locomotion of the individual has a certain degree of freedom which
is char a ct er i sed by a specifi c search st r a t eg y such as a type of ran d o m walk, and is
also subject to other external or internal constraints, such as the environmental context
of the landscape or the physical and psychological conditions of the individual. It is
assumed that a strategy that optimises the search eciency can evolve in response to
these constraints, and the movement is a consequence of the optimisation on random
search.
A seminal work by Viswanathan et al. [27] first studied L´evy-flig ht foraging th r o u g h
the scheme of optimizing random search. In their model, a forager searches for targets
using a random walk with the aforementioned power-law step-size distribut i o n . The
forager will keep mov i ng until a target is ’encountered’, i.e. a target lies within its lim-
ited perception range . The sear ch eciency is defined as the enco u nter r a t e of ta r g et s,
namely th e number o f visited target s per un i t movi n g distance. The model considers
2
Page 2 of 26
http://mc.manuscriptcentral.com/jrsi
Under review for J. R. Soc. Interface
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Mf))F)
'CHAB6RC(<N(@HA(&<:AI(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
MI))F)
'A64CH(*eCDA=CS(6=:(, 4 DbD=>(0 646BA@A4F(
.+'"%-)Qq%#+$%9)
)
S
N
@)$3:7+")2<)5#,/$%0)0'"8+0,)5#,%2S+"+5)79)0-+)<2"'8+")3&)02),0+&)
$)
)))))@)0-+):2S#$8)5#,0'$%+)2<),0+&)$))
)
;-'0)'"+)0-+)5+&+$5+$%#+,)2<)x>)
r3:7+")2<),0+&,)
4'$5,%'&+),#d+)
r3:7+")2<)"+,23"%+)&'0%-+,)
)
h3$)y-'2A)*'G')H3"5'CA)H#'G3$)4#3)+0)'(gA)iE&/:'()4zS9B\#8-0)<2"'8#$8)#$)')p$#0+)('$5,%'&+Aj)'"]#SA)
<2"0-%2:#$8)#$)H23"$'()2<)0-+)*29'().2%#+09)U$0+"<'%+)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
For Review Only
truncated. The actual moving distance l
n
= l l in this case is smaller than
the probabilisic moving distance l and we set L
f
0. There are two situations
of detecting a target: (a) The target is a new target that has not been discovered
before. Then we upda t e S
n
by S
n
S
n
+1andthelocationofthisnewtargetis
memorised by the fora g er . (b) The target is a pr ev i o u sl y visited target. In thi s case
we do not update S
n
. One should n o t e t h a t if this step is the first step the forager
leaves a target, the forag er will ignore th a t target in detect i o n to avoid trapping. If
no target is detected in thi s step, we update L
f
by L
f
L
f
+ l.
2. If the d ecisi on is returning, th e forager will move to one of the previously visited
targets in a straig ht line. Note that the for ag er does not attempt to detect targets i n
areturnstep,whichisanalogoustothe‘blind’phaseinintermittentrandomsearch
[28]. We assume that the forager ca n memorise t h e l ocations of all previously visited
targets and randomly decide on the target of the return phase. In this initial model
we focus on this simple approach to modelling memory and leave more complicated
memory process to future work.
Here we assume that the foraging process starts at a r a n d om tar g et in the land sca pe.
That target can be understood as the base of the forager, and its location is recorded in
the initial memory of the forager. Therefore the forager can have at least one location
to choose in the return phase. We use an indicator function Θ
n
{0, 1} to characterise
the termination condi t i o n of the process. When a step is performed, we update n by
n n +1and checkthevalueofΘ
n
. The process will be continued if Θ
n
=0,andbe
terminated once Θ
n=N
=1whereN denotes the total number of steps upon term i n a t i o n .
We then define the search eciency η as the ratio of the total number of distinct
targets discovered by the forager to the total moving distance upon termination, which
yields
η
S
N
L
N
. (1)
One should note that, besides the subjective returning in the return phase, the forager can
also revisit a previously discovered target if it lies within the forager’s perception range
in the exploration phase. The revisita t i o n during exploration may occur i n two scenario:
(1) the forager taking advantage of the chance proximity of a previously discovered target
to relieve movement constraints (e.g. to rest or supplement energy) prior to reinitiating
5
Page 5 of 26
http://mc.manuscriptcentral.com/jrsi
Under review for J. R. Soc. Interface
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
be another fact o r that aects the for a g i n g eciency. Our study not only sheds new lig ht
on the the understan d i n g of L´evy-flight foraging, but also provides an exp a n ded modelling
framework to study animal m ovement p at ter n s.
2 Model
The foraging takes place in a finite two-dimensional L L squared landscape with per i -
odic boundary conditions (when moving across the boundary the forager will come back
from the other side of th e l an dscape). There are K targets distributed uniformly over
the landsca pe, corresponding to a densi ty = L
2
/K. The forager can detect a ta r g et
within its perception range r
v
. The mean free path of the system is therefore given
by =(2r
v
)
1
, which indicat es the average strai g ht-lin e moving dista n ce of detecting
or ‘encou nterin g’ a target in the landscape. Without loss of g ener a l i ty we set r
v
=1,so
=(1/2). In this paper, we assum e the targets are revisitable, analogous to the case of
non-destructive foraging [27]. The goal of the forager is to explore the lan d sca pe to find
new t a r g et s.
The foraging process is a step-based stochastic process with an exploration-return
mechanism. At each step n, the fo ra g er first decides the type of the step movement:
exploration or return. Let us den ot e the prob ab i li ty of choosing explor at ion by p
n
and
the probability of choosing return by q
n
=1 p
n
. Moreover, we use S
n
to denote the
number of distinct targets discovered by th e forag er up to step n, L
f
to denote the
accumulated moving distance since the forager leaves its last visited target, l
n
to denote
the moving dista nce of step n,andL
n
=
P
l
n
to denote the total moving distance up to
step n. The foragin g movement a t step n is performed as follows (i l l u st r at ed in Figure
1):
1. If t h e decision is explorati on , the forager will perform rand om search in this st ep .
The s tep-si ze l and the turning angle are drawn randomly from the pre-d efi n e d
distribution fun ct i o n s P (l)andP (). During the step movement, the forager will
continuously detect t a rg et s. If a target is detected during the step movement, the
forager will move to t h e target in a straight line and the step movement wi l l be
truncated. The actual moving distance l
n
= l l in this case is smaller than
the probabilisic moving distance l and we set L
f
! 0. There are two situation s
4
ML))F)
$GRB6I('A64CH(*eCDA=CS(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
MM))F)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
,AI6S?#<IA46=CA(
MW))F)
#HA(E6@&!2(G4<@<C<I(
QS+"9)$25+)-',)'$)',,#8$+5),(20)7',+5)2$)$25+)U6)
,(20){)$25+U6)s)$3:7+"|2<|,(20,))
)
)
)
)
e27#(+)$25+,),+$5)'$$23$%+:+$0)7+'%2$,)+S+"9)W):#$30+,@)
Nr25+)U6)
N=&&(#%'/2$)S+",#2$)`+g8g)fgMb)
Ne']#:3:)'S'#('7(+)\',-)&'8+)
NP2(0'8+)
L( J( f( M( g( h( i( j( k( LK( L(
fZ),(20,){)W):#$30+,)
fI@ZZ) fI@Zf) fI@ZI) fI@ZL) fI@ZM) fI@ZW)
!g).2::+"A)Dg)h3,9A)=g)e%h+2V$A)'$5)*g)H3"5'CA)oT-+)D#8)r#8-0)E30@)Q]&+"#+$%+,)<"2:)T"'%C#$8)J(9#$8)J2]+,)
V#0-)6+('9)T2(+"'$0);#"+(+,,)r+0V2"C#$8Ao)U$)&"2%++5#$8,)2<)`*+'(;.rbA)?2:2)4'C+A)U0'(9A).+&0+:7+")IZfLg))
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
MX))F)
1AB<@A(04<CA:94A(26IIF(`102a(
*+,&2$,+B*+K3+,0)&"202%2()
N?(#+$0)`7',+,0'/2$b)
N.+"S+")`:27#(+)$25+b)
N.0'0+(+,,)
Q]':&(+,@)/:+|8+0`bA)/:+|,+0`vbA)"+7220`b)
)
U:&(+:+$0'/2$)
NQ'%-)*!?)-',)')3$#K3+)#5+$/p+")`C+9,g0]0b)
NQ$%'&,3('0+5)#$02)'),#$8(+)"'5#2)&'%C+0@)#5+$/p+")`IB790+,bA)
&'9(2'5)`$BD90+,b)
/:+|8+0`b)
IZfLBZYBII)Zc@LZ@ZZ)
r25+)
D',+)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
MY))F)
)
)
)
)
)
l,+,)*!?),37,9,0+:)
NT"'$,&'"+$0)02)'&&(#%'/2$A)$2)+]&(#%#0)%25+)#$)'&&(#%'/2$)
"+K3#"+5)
Nr+V)&"28"':)0"'$,<+""+5)02)\',-)V-#(+)+]#,/$8)&"28"':)
%2$/$3+,)"3$$#$8)
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
M^))F)
$bA4(@HA(!D4(1AG4<>46BBD=>(
LI)790+,)
r25+)
D',+)
!=1Q)Z)
!=1Q)f)
!=1Q)I)
!=1Q)Z)
!=1Q)f)
!=1Q)I)
!=1Q)fI^)
IWX)790+)&'8+,)
?*?)'%C)
2<=@6C@(/<>>D=>(N<4(,6@6(&9ID=>(
?':&)
=)`D',+b)
?':&)D)
J++5#$8)42%'/2$)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
Mc))F)
,6@6(&9ID=>T($GA=(l9AFR<=F(
}3+,/2$)f@)O2V)02)5+0+%0)%2$0'%0,)7+0V++$)'$#:'(,>)
N.+$5#$8)'$$23$%+:+$0)7+'%2$,>);-'0)'55#/2$'()#$<2":'/2$)
02)#$%(35+>)
N.9$%-"2$#d'/2$>)
}3+,/2$)I@);-'0)#$<2":'/2$),-23(5)V+),02"+)(2%'((9>)
N4288#$8)+S+"9)%2$0'%0>)63"'/2$)2<)%2$0'%0>)
NE$(9)53"#$8)$#8-~:+>)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
WZ))F)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
2<BG4AFFD=>(#467AC@<4DAF(9=:A4(1AF<94CA(
2<=F@46D=@F(
2<=F@46D=@FT(fJ(^ES@AF(1$&m(M(^ES@AF(1!&(
N nV6=@(J(^ES@AF(1!&X(%<m(S<9(W<=o@(AbA=(>A@(L(^ES@AFp(q(F6S(<94(
ABOA::A:(:AbAI<GA4F(
*44<4?O<9=:A:m(C<F@?AeCDA=@(@467AC @<4 S(C <B G4AFFD<=(
#<(F9GG<4@(N9@94A(<=ID=A(:6@6(B6=6>ABA=@d6=6ISFDF(N46BAW<4;(
Wf))F)
#HA(E<9=:A:(l96:46=@('SF@AB(`El'a(
(
=$)2$(#$+)0"'G+%02"9),#:&(#p%'/2$)'(82"#0-:)0-'0))
N <2")$)&2#$0,A)&"2S#5+,)
t E`$b)V2",0)%',+)/:+)%2:&(+]#09`)E`fb)<2")+'%-)#$%2:#$8)&2#$0)b)
t E`fb)V2",0)%',+),&'%+)%2:&(+]#09)
t 83'"'$0++)2$),#:&(#p%'/2$)+""2")`:']#:'()+""2"b)
t 8+$+"'((9)7+m+"),#:&(#p%'/2$)"'0+)
)
)
)
Hg)4#3A)hg)y-'2A)!g).2::+"A)Dg)h3,9A)*g)H3"5'CA)iD23$5+5)}3'5"'$0).9,0+:@)
Qq%#+$0)T"'G+%02"9)?2:&"+,,#2$)2$)0-+)12Aj)U?6Q)fW))
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
WI))F)
U<4B9I6R<=(
1#S+$)')&2#$0)p)',)0-+),0'"0)2<)')0"'G+%02"9),+8:+$0A)<2")+'%-)#$%2:#$8)&2#$0)p
t
)'0)/:+)tA)
p$5)230)#<),-23(5)7+)C+&0)#$)0-+)%3""+$0)0"'G+%02"9),+8:+$0)2")3,+5)',)0-+),0'"0)2<)')
$+V),+8:+$0)
Q]':&(+@)1#S+$)+""2")723$5)I`:+0+",b@)
)
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
start!
d=0!
start!
d=1!
start!
d=3!
p
new!
p
new!
p
new!
start!
d=1!
end/start
new!
WL))F)
El'T(#HA(2<=CAG@(
=)D}.)%2$,#,0,)2<)
N =)723$5#$8)72 ])
N TV2)723$5#$8)(#$+,)
N Q#8-0),#8$#p%'$0)&2#$0,)
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
WM))F)
El'T(#HA(2<=CAG@(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
J2")'),+8:+$0A)0-+"+)%'$)7+)'0):2,0)
M)D}.)+]#,/$8A)+'%-)<2")2$+)
K3'5"'$0)
=%%2"5#$8)02)0-+)&2,#/2$#$8)2<)(
,+
)
'$5)0-+)%2$S+])-3((A)5#k+"+$0)
<2":3('/2$,)2<)(2V+")723$5)'$5)
3&&+")723$5)'&&(9)
E$(9)V-+$)(7a+a37A),-23(5)V+)$++5)
02)&+"<2":)'%03'()5#,0'$%+)
%'(%3('/2$)
!"27(+:@),&'%+)%2:&(+]#09)'$5)
V2",0)%',+)/:+)%2:&(+]#09)
WW))F)
U6F@?El'T(!=(AeCDA=@(6GG4<QDB6R<=(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
l&2$)2%%3""+$%+,)2<)(7a+a37A)
#$,0+'5)2<)&+"<2":#$8)'%03'()
5#,0'$%+)%'(%3('/2$A)V+)0'C+)
')%2$,+"S'/S+)'&&"2'%-)79)
+$5#$8)0-+),+8:+$0)'$5)
,0'"/$8)')$+V)2$+)
=5S'$0'8+,@)(2V)/:+)'$5),&'%+)
%2:&(+]#09)
*#,C,@)%2:&"2:#,#$8)
%2:&"+,,#2$)"'0+)
lb<e<ub ?!
Start point (red)!
Current point (green)!
Processed points (black)!
WX))F)
1AF9I@F(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
?2:&"+,,#2$)"'0+)
)
)
)
)
)
)
=%-#+S+,)(2V+")%2:&(+]#09)'$5)7+m+")%2:&"+,,#2$)"'0+)0-'$)%(',,#%)2•#$+)
'(82"#0-:)'$5)7',+(#$+)2$(#$+)'(82"#0-:,)
Flying fox!
Vehicle!
WY))F)
$GA=(2H6IIA=>AF ((
6+('9B02(+"'$0)v)
6'0'),02"'8+)`V-'0)02),02"+)2")$20b)
.':&(#$8)`:']#:3:)#$<2":'/2$)<2")+$+"89)73%Cb)
?2::3$#%'/2$)`&"#2"#/+,A)<'#"$+,,A)0-"238-&30b)
Q$+"89):'$'8+:+$0)`%2$,3:&/2$A)-'"S+,/$8A)&"+5#%/2$b)
T"'5+2k,>)
e27#(#09):25+(B5"#S+$),':&(#$8)
O2V)02)73#(5)0-+):25+()V#0-230)0-+)5'0')!)'5'&/S+)
:25+(,)
O2V)\+]#7(+)52)0-+,+):25+(,)$++5)02)7+>)
)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
W^))F)
#<('9B("G(
E$82#$8)V2"C)
e25+(#$8):27#(#09))
?2$/$323,)'$5)$+'"B&+"&+03'()0"'%C#$8)7',+5)2$)'%/S#/+,A)
:27#(#09):25+(,)
T"'G+%02"9):+"8#$8R'8#$8))
?2$/$+$0'().%'(+)T"'%C#$8)
r+'"B&+"&+03'():2$#02"#$8)2<)&2,#/2$)'$5)%2$5#/2$))
P+"9)%-'((+$8#$8)9+0)#$0+"+,/$8)"+,+'"%-)&"27(+:)V#0-)"+'()
'&&(#%'/2$)5"#S+",)
?"+'0+,)'8+$0,)<2"):#%"2,+$,#$8)
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
Wc))F)
2<=@4DO9@<4F ((
?.U*E)
D"'$2)h3,9)
!-#(#&&).2::+")
H#'G3$)4#3)
h3$)y-'2)
r'S#$5')h2m+8+)
?-"#,)?"2,,:'$)
!-#()P'(+$%#')
4+,(#+)ES+",)
*2,,)63$8'S+(()
.0+&-+$)D"2,$'$)
;+$)O3)
6'S#5);+,0%2m)
=5':)e%h+2V$)
O'd+()!'""9)
}lT)
!+0+")?2"C+)
l}))
r+#()D+"8:'$$)
lr.;)
.'(#()h'$-+"+)
.'$G'9)H-')
1-3(':)e3"0'd')
43C',)4#)
)
2<II6O<46@<4F(
42$8B0+":)0"'%C#$8)#$)D'0:2$@)4+,,2$,)'$5)E&+$)?-'((+$8+,)F)*'G')H3"5'C)
XZ))F)
!"#$%$&$"'(')'#*&'(+(,-.-#!/(01$,"2#-3-#)(
,45(1676(894:6;(
04D=CDG6I(1AFA64CH('CDA=RF@(
1AFA64CH(.4<9G(/A6:A4(
,DF@4DO9@A:('A=FD=>('SF@ABF(
(
4794:6;rDAAA5<4>(
)
#H6=;()<9(
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.